Abstract
There are more than 72,000 missing in action service members from WWII, many of whom were lost at sea in aircraft. Using high-resolution sidescan sonar and a YOLOv7 model, we present a deep learning approach to expedite the search for these individuals. Our training dataset is the largest aircraft wreck-focused dataset that has been published to date, with 19 unique aircraft wrecks, composed of 290 individual fragments and located across six countries. Our trained model produces an F1 score of 0.74. We tested the model on newly collected data and the model correctly identified three out of four previously unknown aircraft. As the model becomes more accurate and trust in the model is built, we envision a human-in-the-loop approach in which newly collected data are input into the model in the field, the model is run with a confidence threshold of 0.5, and predictions are then checked by the human field team. As trust builds, the confidence threshold can be made higher so that fewer predictions are generated and human review time is minimized.